EP0077089A2 - Device for storing or transmitting transform-coded picture signals and for regaining those picture signals - Google Patents

Abstract

With the usual transformation coding of images, the value ranges of the coefficients generated thereby are enlarged when the individual calculation steps are carried out. As a result, these coefficients contain considerable redundancy, so that the effectiveness of the data reduction in the subsequent quantization is reduced. According to the invention, the calculation steps based on the basic transformation of two values are carried out in such a way that the result of each calculation step is only the original word length, in that a part of the value range of the results is mapped in a different area for each calculation step. An example for the technical implementation of such a mapping to modified coefficients and the inverse transformation of these modified coefficients has been described. In doing so, special quantization curves have to be used, which bring additional advantages.

Description

The invention relates to an arrangement for storing or transmitting and for recovering image signals, in which coefficient values transformed from the image signals obtained by point-by-point scanning of an image are obtained in a transformation arrangement and these are converted into quantized values in a quantizer and the quantized values after storage or transfer decoded in a decoder and the decoded values are converted back in a reverse transformation arrangement into image signals which largely correspond to the original image signals, the transformation arrangement and the reverse transformation arrangement each combining two image signals or decoded values according to the transformation algorithm to form two result values and the result values of different combinations linked step by step.

For storing or transferring images, it is desirable to use as few information units as possible and still display the scanned image as accurately as possible. The number of information units can be reduced if the existing redundancy and possibly also the irrelevance are largely eliminated in an image. It is known, for example from the magazine "IEEE Transactions on Computers", Vol. Com-19, No. 1, February 1971, pages 50 to 62 or from the book by Pratt "Digital Image Processing", John Wiley & Sons 1978, Pages 232 to 278, for this reduction in the number of information units, ie for a data compression transform coding method with subsequent quantization. When quantizing The transformation coefficient is usually a non-linear characteristic curve that is based on the calculation of statistical averages.

Wenn nun beide Bildsignale A und B den maximalen Wert hatten, hat der Koeffizient F (1) den doppelten Wert, d.h. es ist eine Verdoppelung des Wertebereiches aufgetreten. Bei der Differenzbildung entsteht durch das Vorzeichen eine Verdoppelung des Wertebereiches. Da bei der Transformation eines vollständigen Unterbildes die Bildsignale in mehreren Stufen verarbeitet werden, wobei jeweils eine Verdoppelung des Wertebereiches auftritt, ist der Wertebereich der transformierten Signale größer als der Wertebereich der Bildsignale, so daß die Datenreduktion durch die nachfolgende Quantisierung dadurch zumindest stark verringert wird. Eine einfache Verkleinerung des Wertebereiches der transformierten Signale oder Koeffizienten auf den ursprünglichen Wertebereich durch eine entsprechende Normierung der Koeffizienten würde zu unerwünschten Rundungsfehlern führen.When transforming image signals, however, the value range of the individual transformed signals increases, as is briefly indicated by means of a Walsh-Hadamard transformation of two image signals A and B. The transformation of these two pixels yields the two coefficients

If both image signals A and B had the maximum value, the coefficient F (1) has twice the value, ie the value range has been doubled. When the difference is formed, the sign doubles the value range. Since the image signals are processed in several stages during the transformation of a complete sub-image, whereby the value range is doubled in each case, the value range of the transformed signals is larger than the value range of the image signals, so that the data reduction due to the subsequent quantization is at least greatly reduced. A simple reduction in the range of values of the transformed signals or coefficients to the original range of values by appropriate standardization of the coefficients would lead to undesired rounding errors.

The object of the invention is to provide an arrangement of the type mentioned at the outset which avoids an enlargement of the value range of the values to be stored or of the values to be transferred compared to the original value range of the image signals, at the same time being a favorable one Quantization is possible. This object is achieved according to the invention in that the transformation arrangement has at least one additional logic element which receives the carry signal of the one result value which arises with each logic operation and the sign signal of the other result value and generates an auxiliary signal at the output which is fed to the one result value as the most significant bit, the least significant bit of this result value and the carry signal and the sign signal of the result values being suppressed for storage or transmission.

The invention uses i.a. the fact that the results of summing and subtracting two numbers must always be both even or both odd so that the last bit of one of the two result values can be omitted without losing information. The combination of carry and sign of the result values does not actually result in any ambiguity, but the result values can be clearly converted back during the back transformation, as will be explained later. In this way, since the result values only have the same value range as the linked values for each link, the values to be quantized have the same value range as the original image signals. The occurrence of redundancy generated by the transformation itself is thus avoided, so that an optimal data compression can be achieved together with a quantization.

With some transformation algorithms, especially with the already mentioned Walsh-Hadamard transformation, only very simple links occur. An embodiment of the invention for this is characterized in that the transformation arrangement comprises the two result values of each combination of the sum and the difference between two values Forms that the additional logic element is an exclusive OR gate and that the reverse transformation arrangement each links a first decoded value containing the least significant bit with a second decoded value still containing the auxiliary signal and thereby the bits of the first value by one digit shifts to the lower value and assigns the value "0" to the previous bit position of highest value and processes the bit of the lowest value as a carry signal for the lowest position in the summation. This results in a very simple design of the additional linking element, which is independent of the value range of the image signals, and the individual result values are also formed in a simple manner during the back-transformation, without the carry-over and the sign having to be explicitly converted back.

Since the combination of the carry and the sign of the result values represents, mathematically speaking, another representation of the result values in the display space, the characteristic curve of the quantizer or decoder must expediently also be adapted to this. A further embodiment of the invention is characterized in that the quantizer works only in the value range of the original image signals with quantization characteristics which are symmetrical or, in many cases, symmetrical with respect to coefficient values that are equidistant from one another. This also results in better maintenance of pronounced edge structures in the image, as will also be explained later.

• Fig. 1 ein grobes Blockschaltbild einer gesamten Anordnung zur Transformation und Rücktransformation von Bildsignalen,
• Fig. 2 ein Blockschaltbild einer Transformationsanordnung oder Rücktransformationsanordnung,
• Fig. 3 schematisch die Abbildung von Teilen des Wertebereiches der Ergebniswerte auf einen anderen Wertebereich,
• Fig. 4a eine Anordnung zur Verknüpfung von zwei Werten sowie ein zusätzliches Verknüpfungselement zur Verwendung in der Transformationsanordnung,
• Fig. 4b eine Anordnung zur Verarbeitung von zwei transformierten Werten zur Verwendung in der Rücktransformationsanordnung,
• Fig. 5a und b zwei Quantisierkennlinien für zwei verschiedene Koeffizienten.

An embodiment of an arrangement according to the invention is explained in more detail below with reference to the drawing. Show it

• 1 is a rough block diagram of an entire arrangement for transforming and re-transforming image signals,
• 2 shows a block diagram of a transformation arrangement or reverse transformation arrangement,
• 3 schematically shows the mapping of parts of the value range of the result values to another value range,
• 4a shows an arrangement for linking two values and an additional linking element for use in the transformation arrangement,
• 4b shows an arrangement for processing two transformed values for use in the reverse transformation arrangement,
• 5a and b two quantization curves for two different coefficients.

In Fig. 1, the image signals are supplied via the input 1. These image signals can be supplied directly from the image scanner, such as a television recording device, but it is generally more expedient to temporarily store the image signals of an image in a memory, since a larger image is generally transformed and transmitted in successive, usually square sub-images. This possibility is not shown, however, since it is not relevant to the invention.

The image signals supplied via line 1 are transformed in transformation arrangement 2 in accordance with the desired transformation algorithm under control of control unit 6, and the transformed values or coefficients are supplied via line 3 to quantizer 4, which is also controlled by control unit 6 becomes. The quantized values are fed to the transmission link 10. Instead of the transmission link in particular, a memory can also be used.

The values transmitted or read out from the memory are fed on the receiving side to a decoder 14, where the individual quantized values are assigned an average coefficient value which belongs to the quantization level indicated by this value. The decoded coefficient values are fed via line 13 to the reverse transformation arrangement 12, so that essentially the original image signals supplied to the input 1 appear at the output 15.

The decoder 14 and the inverse transformation arrangement 12 are controlled by the control unit 16, which also contains the characteristic curve memory for the assignment of input values to coefficient values in the decoder 14, just as the control unit 6 on the transmission side contains the characteristic curve memory for the quantizer 4.

The general structure of a transformation arrangement 2, a quantizer 4, a decoder 14 and a reverse transformation arrangement 12 is known in principle, the structure of the transformation arrangements also being able to be different for a given transformation algorithm.

mit

For the further explanation, the mathematical background should first be indicated, using a Walsh-Hadamard transformation. As a basic step, a 2-point transformation of two pixels A and B is carried out:

With

The resulting transformed signals or coefficients are without standardization, as already stated at the beginning

These coefficients can therefore easily be determined with the help of conventional technical arithmetic modules, which are only constructed for processing two values each.

A Walsh-Hadamard transformation for an entire image or sub-image of 2n (n = 2 ^N , N = 1,2, ...) pixels can be derived step by step from the specified basic step in the following way

Thus, two coefficients formed from different basic transformation steps are processed further with such a basic transformation step, etc., until finally all pixels of the named picture or sub-picture are taken into account in each transformed signal. The transformation therefore takes place in several stages. There are of course also arrangements parallel to the bar, in which the matrix multiplication is carried out quasi in parallel, but this requires an extraordinarily great effort, which is justified only with very extreme speed requirements. The transformations that are normally carried out in practice can always be traced back to the step-by-step processing of two values.

An arrangement for carrying out a Walsh-Hadamard transformation in this way is shown in FIG. 2. Every second signal of the signals supplied at the input 21 is temporarily stored in the buffer 24, so that two successive signal values are present at the input of the computing unit 22. In the Walsh-Hadamard transformation, the computing unit 22 forms the sum and the difference of these two values and outputs the results at the two outputs 23 and 25 designated accordingly. These initial values correspond to the coefficients F (0) and F (1) given in equation (3), which are given in FIG. 2.

The first value of the output values at the output 23 is temporarily stored in the buffer memory 28 and then fed to the one input of the computing unit 26, while the second output value of the output 23 is fed directly to the other input of the computing unit 26. The corresponding process is carried out with the output values at the output 25, of which the first value is temporarily stored in the memory 32 and then fed in parallel to the computing unit 30 together with the respective second value. The computing unit 26 and 30 are constructed in exactly the same way as the computing unit 22 and also the following computing units 34, 38, 42 and 46.

Output values then appear at the outputs of the computing units 26 and 30, which correspond to the coefficients of a sub-image consisting of 2x2 pixels. These output values are also alternately fed via the intermediate memories 36, 40, 44 or 48 or directly to the computing units 34, 38, 42 and 46. These computing units generate the coefficients F "(0), F" (1) ... of a sub-picture with 4x2 pixels at the outputs.

This arrangement can be continued as desired that correspondingly large sub-images are transformed and the coefficient values are output in parallel. However, since the arithmetic unit 22 only outputs two output signals in parallel for every second image signal supplied, and this applies to all subsequent arithmetic units, it may be expedient to temporarily store all the output signals of the arithmetic unit 22 for a sub-image of a predetermined size and, after processing all the image signals of this sub-image, the input of the arithmetic unit 22 switch to the output of this buffer and repeat this process for each stage according to the specified sub-picture size. For such an arrangement, which works according to the so-called "pipeline principle", a special control with a special address generator is necessary, but there is the advantage that only a single computing unit is required.

In order to prevent the value range of the output signals from doubling in each computing unit or at each stage of the processing compared to the value range of the input signals, the result values are now mapped differently, as will be explained with reference to FIG. 3. The sloping coordinates A and B indicate the possible input values of an arithmetic unit. The Walsh-Hadamard coefficients F (0) and F (1) formed from this are given in the perpendicular coordinate system. From this it can be seen that the coefficients F (0) and F (1) each have a value range of 2 G if the value range of the signals A and B supplied is equal to G. The enlargement of the value range arises from the fact that a carry-over can occur when the sum is formed and a negative sign can occur when the difference is formed. On the other hand, it can be seen from FIG. 3 that the value combinations of the coefficients F (0) and F (1) indicated by crosses are formed by all combinations of the input values A and B. will not include all possible value combinations of the coefficients within their entire value range. This means that the coefficients contain redundancy generated by the transformation itself. This is determined in more detail below. From equation (3) it follows that the coefficients F (0) and F (1) are either both even or both are odd. When these values are represented in dual form, they always match in the least significant bit, so that only one coefficient has to take this into account.

Diese letztere Beziehung beschreibt die Anordnung der tatsächlich auftretenden, durch Kreuze bezeichneten Wertekombinationen der beiden Koeffizienten F (0) und F (1) in Form eines auf der Spitze stehenden Quadrates, so daß von dem gestrichelt gezeichneten äußeren Quadrat, das die Gesamtzahl der darstellbaren Wertekombinationen umfaßt, die Bereiche in den Ecken nicht besetzt sind, so daß die Anzahl der darstellbaren Wertekombinationen einen Bereich umfassen, der einer Verdoppelung des Bereiches der tatsächlich auftretenden Wertekombinationen der Koeffizienten F (0) und F (1) entspricht. Eine weitere Verdoppelung ist dadurch gegeben, daß im darstellbaren Wertebereich auch Kombinationen von Koeffizienten vorhanden sind, von denen der eine gerade und der andere ungerade ist, während bei den tatsächlich auftretenden Wertekombinationen gemäß der Gleichung (3) die Koeffizienten jeweils nur beide gerade oder beide ungerade sein können.It also follows from FIG. 3 that the value combinations of the coefficients F (0) and F (1) indicated by crosses fulfill an uncertainty principle of the following form:

This latter relationship describes the arrangement of the actually occurring combinations of values, denoted by crosses, of the two coefficients F (0) and F (1) in the form of a square standing on the top, so that the dashed outer square represents the total number of representable combinations of values includes, the areas in the corners are not occupied, so that the number of combinations of values that can be represented include a range that corresponds to a doubling of the range of actually occurring value combinations of the coefficients F (0) and F (1). A further doubling is given by the fact that combinations of coefficients are also present in the range of values that can be represented, one of which is even and the other is odd, while in the case of the actually occurring value combinations according to equation (3), the coefficients in each case only both even or both odd could be.

mit G = Anzahl der Werte der Variablen A bzw. B.This restriction of the combinations of values of the two Coefficients can now be used to form transformed coefficients whose value range is not larger than that of the input values A and B supplied. For this purpose, at least some of the coefficients initially formed are mapped in another area of the event field shown in FIG. 3. Various images are possible. Each of these images is based on a division of the original event field shown in FIG. 3 into four areas I, II, III and IV, which is carried out by the transfer of the sum coefficient F (0) and by the sign of the difference coefficient F. (1) can be determined. It is assumed that the difference A - B for the difference coefficient F (1) is represented in two's complement:

with G = number of values of variables A or B.

As a result, positive differences are represented by 1 bit with the value "1" in the most significant position, which is referred to as the sign bit VZ, and negative differences contain the value "0" in the most significant bit position. With the sum coefficient F (0), the bit position of the highest value indicates the carry Ü. By combining the carry bit Ü and the sign bit VZ of the coefficients F (0) and F (1), different mappings can be realized, so that the entire range of values of the two coefficients to be displayed is halved together. By additional use of the first condition, i.e. The property that only both coefficients can be even or odd results in modified coefficients whose value range is equal to that of the values A and B supplied. With such a mapping, there is no loss of information, so that the pixels can be reconstructed exactly by means of a corresponding reverse transformation.

3 shows an imaging method in which area I is shifted to the right above area IV and area III is shifted to the right below area II. The resulting modified coefficients F ^* (O) and F ^* (1) are thus formed in accordance with Table 1 below, the two-complement representation being used as the basis for the difference formation:

In contrast to FIG. 3, a number G = 2 ⁴ = 16 values of the input variables A and B is assumed here. The shift is thus formed by an exclusive NOR operation of the carry bit Ü and the sign bit VZ.

A circuit arrangement which realizes such a mapping of the coefficients and which can be used for each of the computing units 22, 26, 30 etc. in FIG. 2 is shown in FIG. 4a. The two signals A and B supplied are assumed to be four-bit dual words, which is indicated by the block with four boxes in the signal path. These two signals are fed to both an adding unit 60 and a subtracting unit 62, the subtracting unit 62 being supplied with a carry signal with the value "1" in order to obtain the corresponding difference representation in two's complement. The output values of the two units 60 and 62 represent the coefficients F (0) and F (1), which are represented by dual words with five bits. With the coefficient F (0), the bit Ü indicates the Instead of the highest value, the carry and for the coefficient F (1) the bit VZ at the position of the highest value, the sign.

These two bits are now fed to the inputs of an exclusive NOR gate or equivalent gate 64, and the output signal of this gate replaces the carry bit Ü at the coefficient F (0). Furthermore, the cross-valued bit of the lowest value of this coefficient is omitted since the corresponding bit of the other coefficient F (1) has the same value. This results in the modified coefficient F ^* (0), which is only four bits long, ie contains the same number of information units as the supplied values A and B. With the coefficient F (1), the sign bit is not continued because it is already indirectly in the most significant bit of the modified coefficient F ^* (0) is contained, but only the last four bits are passed on as a modified coefficient F ^* (1). This coefficient also has the same number of information units as the supplied values A and B. Overall, the coefficients F ^* (0) and F ^* (1) according to Table 1 thus result, which contain the same number of information units as the supplied values A and B. If the computing units 22, 26, 30 etc. in FIG. 2 are realized by an arrangement according to FIG. 4a, it is clear that even with a longer cascading of computing units for the transformation of a larger sub-picture there is no increase in the word length of the coefficients arises.

The reverse transformation unit 12 in FIG. 1 is constructed similarly to the transformation unit 2; when using the Walsh-Hadamard transformation, the two arrangements are even identical. Finally, by step-by-step processing of two coefficients, apart from the quantization errors, the original image data is generated. When using the modified coefficients that are generated by the arrangement according to FIG. 4a, however, a modified inverse transformation is also necessary.

First, the mathematical background of the back transformation of the modified coefficients will be explained. When using the Walsh-Hadamard transformation, the sum and the difference of two coefficients or, in the following steps, two intermediate values are formed in the same way as for the transformation. Taking into account the modified coefficients, the calculation for the back-transformed values A and B, which therefore do not directly represent the back-transformed image signals here, but the result values for each intermediate step, results in the calculation shown in Table 2 below:

The values that are linked in the individual areas are shown in Table 1. The factor 1/2 for the total coefficient F ^* (0) results from the fact that the The least significant bit is omitted. In the case of the difference coefficient F ^* (1), the factor 1/2 also results from the fact that the least significant bit is separated and processed differently from the other bits, as subsequently based on a technical implementation in the form of a computing unit for performing the arithmetic operation described above is explained. It can be seen from Table 2 that with this inverse transformation, in which the subtraction in two's complement is again used by adding the value 2 ⁴ to the difference, the original values A and B are immediately recovered, only with coefficients from certain ranges according to FIG 3 a carry occurs in the form of the summand 2 ⁴ , which can therefore be simply eliminated by limiting the output values to the last four bits of the data words produced during processing.

Such a computing unit is shown in Fig. 4b. This also contains an adding unit 68 and a subtracting unit 66, to each of which two decoded coefficients or intermediate values are fed in parallel, of which the one coefficient, which is designated here with F ^* (0), was last created during the transformation from an addition and the other coefficient, which is designated here with F ^** (1), was last created during the transformation from a subtraction. This latter coefficient is obtained from the coefficient F ^* (1) in that it has been extended by one bit in the most significant position, this additional bit having the value "0". The four most significant bits together with the first bits of the coefficient F ^* (0) are now fed from the coefficient F * ^* (1) to both the subtracting unit 66 and the adding unit 68, the latter as the carry signal the least significant bit of the coefficient F ^{* *} (1) receives while the subtracting unit 66 constantly receives a signal with the value "1" as a carry signal. It could initially be assumed that the least significant bit of the modified coefficient F ^** (1) would also have to be returned to the modified coefficient F ^* (0). However, since two bits of the same value are then added in the last digit during the addition, a carry would occur if these have the value "1". For this reason, this bit is therefore only supplied to the adder unit 68 as a carry signal. In contrast, the value of this bit is of no significance for the subtracting unit 66 and is not processed there.

The most significant bit, which can contain a carry, is separated from the output signals of units 66 and 68 and is not further processed, since the four lower bits immediately indicate the value sought, as has already been explained with reference to Table 2. In this way it is thus possible to avoid increasing the word width even during the back transformation, the two units 66 and 68 each only having to process the number of bits which also contain the modified coefficients supplied.

Since a complete Walsh-Hadamard transformation on a larger number of pixels can be derived step by step from the 2-point transformation according to equation (4), each result value can also be interpreted as a modified coefficient of the next stage and in exactly the same way during the reverse transformation Be processed further. Exceeding the word length is thus avoided both during the transformation and during the reverse transformation, so that the computing units only have to be designed for the smallest possible word length. Furthermore, since the coefficients to be finally quantized have the same word length as that of the supplied image signals, there is a significant simplification and a better effectiveness of the quantization and thus also the data reduction.

In addition, there is also an additional gain in the quantization, as a result of which a more precise representation of the quantized values can be achieved with comparable effort. Large values of the transformation coefficients of each individual transformation stage are precisely represented by quantization characteristics, which are specially adapted to the developed method, so that even pronounced edge structures are better reproduced in the image. than with conventional methods. This will be explained in more detail below using two examples.

First, consider a 2-point basic transformation of the two pixels A and B, which can each take one of a number G values. This transformation leads to the two coefficients F (0) and F (1) given in equation (3). In the previously common methods, the sum coefficient F (0) is transmitted unchanged, while the difference coefficient F (1), which can also take negative values, is quantized for data compression by a nonlinear characteristic curve that is symmetrical to the zero point and then by means of a smaller number of information units is transferred or saved. The non-linearity is chosen so that frequently occurring small differences are finely quantized, while less frequently occurring large difference values are correspondingly coarser.

• Kleine negative Differenzen ergeben große positive Differenzen, große negative Differenzen ergeben kleine positive Differenzen.

When using the described modified coefficients, however, the sign of the difference coefficient F (1) is separated and not transmitted. The deletion of the sign results in the following figure when forming the difference by displaying in two's complement:

• Small negative differences result in large positive differences, large negative differences result in small positive differences.

The condition that small positive differences as well as small negative differences have to be quantized in fine steps because of their greater probability of occurrence leads to a symmetrical quantization characteristic curve, which is shown in FIG. 5a. This quantization curve also gives some additional advantages.

The first advantage arises from the fact that the quantizer only has to process positive values for the difference coefficient. When using a strongly non-linear quantization with a very fine-level quantization in the range of small differences in terms of amount, the number of reduced information units is essentially given by the fine-level quantization. For the quantization curve according to Fig. 5a, however, due to the required symmetry, an almost double-exact quantization in half of the original range is required, so that in this case there is only a small gain for data compression with a constant quantization error if this occurs with the corresponding quantization of the already mentioned characteristic curve symmetrical to the zero point is compared. In the other case of linear quantization, however, the modified coefficients result in a gain of a factor of 2, since the same number of quantization stages is used only in half of the original range. This results in a profit with a factor between 1 and 2 for the modified coefficients.

A second advantage results from the fact that, due to the symmetrical, non-linear quantization characteristic, large values of the difference coefficient are more precisely shown advertises so that even pronounced edge structures, ie steep transitions in the image, are better reproduced than with conventional methods. As a second example, consider a Walsh-Hadamard transformation of a sub-image with four pixels A, B, C and D, which can be carried out by applying the basic transformation twice according to equations (1) and (4). If the modified values according to Table 1 are formed each time the basic transformation is carried out, which does not result in an increase in the word length, the following four coefficients result:

The sum coefficient F ^* (0) is in turn transmitted unchanged. With the coefficient F ^* (1) it must be taken into account that in the first transformation step the most significant bit of the sums A + B or C + D representing the carry is changed by the equivalence function according to Table 1 and the bit of the lowest value is separated, what needs to be considered in the subsequent subtraction. The actual subtraction in the second transformation step then in turn requires a symmetrical characteristic curve according to FIG. 5a, so that the double-symmetrical characteristic curve shown in FIG. 5b is produced overall. The same quantization characteristic results for the coefficient F * (2). The coefficient F * (3), which arises from a two-fold difference formation, in turn only requires the single-symmetrical characteristic curve according to FIG. 5a. In this example too, the formation of the modified coefficients results in a gain for the quantizer, since the quantizer only has to process coefficients with a value range equal to that of the original image signals. In addition pronounced edge structures in the image are also better reproduced due to the symmetry of the quantization characteristic.

This also applies to the transformation of larger sub-images with correspondingly more pixels, which requires correspondingly more stages with the implementation of the basic transformation. The resulting coefficients are then quantized by means of symmetrical or multiple-symmetrical nonlinear quantization characteristics, which are to be adapted in accordance with the two examples described to the sequence of the individual transformation steps according to Table 1.

Claims (3)

1. Arrangement for storing or transferring and for recovering image signals, in which coefficient values transformed from the image signals obtained by point-by-point sampling of an image are obtained in a transformation arrangement and these are converted into quantized values in a quantizer and the quantized values after storage or transfer in one Decoders are decoded and the decoded values are converted back in a reverse transformation arrangement into image signals which largely correspond to the original image signals, the transformation arrangement and the reverse transformation arrangement in each case combining two image signals or decoded values according to the transformation algorithm to form two result values and the result values of different combinations in each case stepwise further linked, characterized in that the transformation arrangement (2) has at least one additional linking element (64) which carries the carry signal receives a result value with each link and the sign signal of the other result value and generates an auxiliary signal at the output, as a result value is added as the most significant bit, the least significant bit ie the result value as well as the carry signal and the local signal of the result values for storage suppress transmission: become. 2. dilution according to claim 1, characterized in that transformation arrangement (2) the two result values of each link from the sum and the difference two values forms that the additional logic element (64) is an exclusive OR gate and that the reverse transformation arrangement (12) links a first decoded value, which still contains the least significant bit, with a second decoded value, which still contains the auxiliary signal, and thereby shifts the bits of the first value by one position to the lower value and assigns the value "0" to the previous bit position of highest value and processes the bit of the lowest value as a carry signal for the lowest position in the summation. 3. Arrangement according to claim 1 or 2, characterized in that the quantizer works only in the range of values of the original image signals with quantization characteristics which are symmetrical or, in many cases, symmetrical with respect to coefficient values lying at the same distance from one another.

Images